Assignment 3
Points:
- Object detection: 25 points
- Drone tracking: 50 points
- Vehicle counting: 25 points
Drone follow me using Kalman Filters
In this assignment, the object class is car the ability to track these objects will be demonstrated using Kalman Filters.
Task 1: Object Detection
Perform object detection on the following video.
Split the videos into frames and use an object detector of your choice, in a framework of your choice to detect the car. Produce a video that has the bounding box and its centroid superposed on the original video.
The video can be downloaded using yt-dlp
Task 2: Kalman Filter
Use the filterpy library to implement Kalman filters that will track the the vehicle in the video. You will need to use the detections from the previous task to initialize and run the Kalman filter. Note that the drone is crashing on a tree at the end of the video and stops.
Since we have no ground truth, we will not be able to produce an estimate of how well the filter is doing in tracking the object and “covering the gaps” when the object detector missed the vehicle due to occlusions. The only information we have is the measurements from the object detector and so the only way we can looking at filter performance is to compare the sequence of measurements (the centroid of the bounding box) with the sequence of estimates produced by the Kalman filter. The so called innovation sequence is the difference between the measurements and the estimated (predicted) measurement. Since the measurement vector is 2D, the innovation sequence is also 2D and you can store its magnitude (norm) at each time step and plot it into a time series plot.
Task 3: Count Vehicular Traffic
You are now tasked to monitor and record the number of vehicles that pass through this highway per direction. You can assume that the video is taken from a stationary camera. Implement your approach that report the two counts (one for each direction) as a function of time (count should increase from 0 to some number per direction at the end of the video and should be normalized by T the duration of the video).